Original article link: https://mp.weixin.qq.com/s/qFwTYorpCMSZ6M1o3CwX_w
To address the insufficient precision of deep learning algorithms in atmospheric environment forecast applications, Feng Lin, an associate researcher from the Institute of Urban Meteorology, CMA, Beijing (IUM), in partnership with researchers from the Beijing Meteorological Observatory and Nanjing University of Information Science & Technology, has developed an atmospheric environment forecast model named PPN (Air Pollution-Predicting Net), tailored to forecast the spatiotemporal distribution of regional PM2.5 concentrations.
Model Framework
From a forecast experiment conducted in January 2022 for the Beijing-Tianjin-Hebei region, the outcomes are as follows: The PPN model can generate PM2.5 concentration forecasts for the entire region spanning 0-72 hours in just 1 second (with a horizontal grid resolution of 9 km), and its precision notably exceeds the forecasts yielded by the traditional CTM-WRF-Chem. Additionally, the original goal of the model design has been achieved: the spin-up phase enhances the initial field, resulting in a significant improvement in forecast precision during the 0-24 hours. The utilization of a weighted loss function in the process of training has also decreased forecast errors for extreme events.
This model can provide support for regional and urban pollution prevention, control, and rapid warnings. Given sufficient computing resources, the model can additionally broaden its scope of forecast and encompass a wider variety of pollutant categories.
Article link:
Qiu, Y., J. Feng, Z. Zhang, X. Zhao, Z. Li, Z. Ma, R. Liu, and J. Zhu, 2023: Regional aerosol forecasts based on deep learning and numerical weather prediction. npj Clim. Atmos. Sci.,6, 71, https://doi.org/10.1038/s41612-023-00397-0.
Related link：
Feng, J., Y. Li, Y. Qiu, and F. Zhu, 2023: Capturing synoptic-scale variations in surface aerosol pollution using deep learning with meteorological data. Atmos. Chem. Phys., 23, 375–388, https://doi.org/10.5194/acp-23-375-2023.
